Precise instructions and constraints

Here's where most prompting frustration comes from: the model does what you wrote, not what you meant. Three habits kill 80% of that.

1. Say what to do, not what to avoid

"Don't be long-winded" is a lousy target — where's the line? "No corporate jargon" — fine, then what instead? Tell the model what to do:

• ❌ "Don't make it complicated" → ✅ "Sentences under 15 words, like you're explaining it to a teenager"
• ❌ "No fluff" → ✅ "Every sentence carries a fact or an action"

2. Hard constraints

Numbers beat adjectives. "Short" means something different to everyone; "under 100 words" means the same to all. Constrain length, the number of bullets, the vocabulary ("no buzzwords"), the shape ("exactly 3 paragraphs").

3. A success criterion

Tell the model how to check its own work: "A good answer: someone with zero experience gets it on the first read and can repeat the steps". The model measures itself against your criterion — and the quality jump is obvious.

Bonus: put the important part last

Models pay the most attention to the start and the end of a prompt. Repeat the one requirement that matters as your final line: "Reminder: 100 words maximum".

Why it takes you so literally

Remember from part 1 of the track: a language model doesn't read minds — it continues text. Whatever you didn't write doesn't exist for it; whatever you did write is a signal. Hence the pink elephant trap: tell someone "don't think about a pink elephant" and guess what they think about. Same with the model — "don't mention competitors" puts the word "competitors" into the text and statistically raises the odds they show up. The positive version — "talk only about our product" — is far more reliable, because it steers instead of forbidding.

Before and after: three real instructions

• ❌ "Make the meeting agenda decent" → ✅ "A 30-minute agenda: 3 items, each with an owner and the decision we want. Most contentious one first"
• ❌ "Write a post, but don't make it salesy" → ✅ "A first-person post: one customer story, one number for the result, one question to readers at the end. No 'buy now' lines"
• ❌ "Make the job ad appealing" → ✅ "Job ad: 5 'what you'll do' bullets starting with verbs, 3 'what you get' bullets with numbers, requirements — only the deal-breakers, 6 max"

Same pattern every time: on the left, judgements ("decent", "appealing") the model interprets however the dice land. On the right, things you can measure.

The mistakes that break even good prompts

• Requirements that fight each other: "detailed but brief" — the model picks one, and not necessarily yours. Decide for it: "brief on each point, but no more than seven points."
• Fifteen rules at once: by rule ten the model is losing rule one — pick the 3–4 that matter and handle the rest through iteration (that's the next lesson).
• Unspoken expectations: you "obviously meant" a formal tone but never wrote it. The model isn't psychic, it's statistics.

A criterion at work: watch the text change

The task: "explain two-factor authentication to our customers." With no criterion the model gives you a perfectly correct encyclopaedia entry — technically flawless, unreadable. Add one line: "Criterion: my 62-year-old mum understands it first time and can switch it on herself" — and the text rebuilds itself. The jargon disappears, steps appear, and so does an example with the code from a text message.

A criterion works like a quality gate on a production line: the model runs its own answer past your check before you ever see it. This is the most underrated trick in the lesson — one line, and half your iterations never happen.

Roll the whole lesson into one four-second check before you hit send: "any bans instead of instructions? any judgements instead of numbers? anything contradicting anything? is there a success criterion?" Four seconds, and your prompt beats most. It pays off double in prompts you plan to save and reuse: a flaw in a one-off costs you one iteration, a flaw in a template repeats every week until you catch it.

Do this now

Dig one "don't do X" and one judgement word ("better", "simpler") out of your recent prompts. Rewrite them: the ban becomes a positive instruction, the judgement becomes a number or a checkable rule. These are the two highest-return reflexes in prompting.